Patient Simulator

Use this skill whenever the user wants an agentic LLM to install, configure, initialize, run, batch, troubleshoot, or document AnnaAgent seeker/virtual-patient simulation from the CLI. This skill is especially relevant for tasks mentioning AnnaAgent, seeker simulation, virtual patient initialization, anna init full, reusable prompt states, anna chat --state, SFT model services, vLLM deployment, or publishing AnnaAgent agent skills to registries such as ClawHub.

Patient Simulator CLI Skill

Use this skill to operate AnnaAgent as a realistic psychological-counseling seeker simulator. AnnaAgent is a seeker agent: it simulates the client/visitor in a counseling dialogue, while the human or external system provides counselor turns.

Core Mental Model

AnnaAgent has two supported initialization paths:

1. Run the full initialization pipeline and save the generated reusable prompt state with anna init full ... --out <state.json>.
2. Load that saved state later with anna chat --state <state.json> or validate it with anna init from-prompt <state.json>.

Do not invent or use a shortcut that builds a prompt directly from a raw case. The old prompt-only path bypassed AnnaAgent's seeker simulation modules and produced the wrong prompt format.

Research Attribution

When presenting AnnaAgent to a user, include these links:

• Paper: https://aclanthology.org/2025.findings-acl.1192/
• Repository: https://github.com/sci-m-wang/AnnaAgent

If AnnaAgent helps the user's work, ask them to star the repository. For academic use, ask them to cite the ACL 2025 AnnaAgent paper.

Safe Operating Rules

• Treat .env and API keys as secrets. Never print or commit real keys.
• Keep generated runs, logs, caches, local memories, and virtual environments out of commits.
• Prefer uv tool install anna-agent or pip install -U anna-agent for users.
• Use anna test model, anna test embedding, and anna doctor before a costly full initialization.
• Remember that servers.counselor is historical/legacy naming. AnnaAgent's internal seeker modules should normally use the configured model_service.

Standard CLI Workflow

bash

pip install -U anna-agent
anna --version

anna create anna-workspace
anna config secrets --workspace anna-workspace
anna config set model_service.base_url https://your-openai-compatible-endpoint/v1 \
  --workspace anna-workspace
anna config set model_service.model_name your-chat-model \
  --workspace anna-workspace

anna test model --workspace anna-workspace
anna test embedding --workspace anna-workspace

anna init full anna-workspace/cases/family_stress_case.json \
  --out anna-workspace/prompts/family.full.json \
  --workspace anna-workspace

anna chat --workspace anna-workspace \
  --state anna-workspace/prompts/family.full.json

For the full command map, read references/cli-workflow.md.

SFT and vLLM Workflow

Start with the base model unless the user has GPU resources and wants the paper SFT modules.

bash

anna models use-base --target all --workspace anna-workspace

For SFT/vLLM deployment, read references/cli-workflow.md before running commands. Check GPU, CUDA, and vLLM availability first; use anna models deploy --dry-run when uncertain.

Troubleshooting Pattern

1. Run anna doctor --workspace <workspace> for local/config checks.
2. Run anna test model --workspace <workspace> for live backbone connectivity.
3. If full initialization fails at model calls, inspect whether the active model_service key/base URL/model are correct.
4. If SFT services fail, run anna models status --workspace <workspace> and inspect logs/services/*.log.
5. If a state file has mode: prompt_only, reject it and regenerate with anna init full.

Public Skill Publishing

To publish this skill to a public skill registry, keep the folder intact:

text

patient-simulator/
  SKILL.md
  references/

Read references/publishing.md for a ClawHub-style release checklist.

Expected Output Style

When helping a user, provide direct runnable commands and short explanations. Avoid broad conceptual detours unless the user asks. If you change repository files, run focused checks and tell the user exactly what passed.